Estimating how much carbon a forest holds usually comes down to an equation that maps measurable tree geometry, such as height, crown size and laser-return density, to dry biomass. Two habits dominate. One fits a single pooled equation for every species. The other fits a separate model for each species, which the evidence favors when data allow: in a Beijing afforestation stand, species-specific random forests beat pooled ones and reached $R^2 = 0.89$ (Forest emissions reduction assessment from airborne LiDAR data using multiple machine learning approaches). Separate models, however, multiply parameters by the number of species. That is a poor trade when a rare species has a few dozen measured trees. It also hides what actually differs between species. A large plot study found that tree size (ontogeny) drives allometric error far more than evolutionary relatedness (Allometric Equations for Aboveground Biomass Estimation in Natural Forest Trees: Generalized or Species-Specific?). That finding hints that most of a biomass curve may be shared, with species differing only in a few places.
We propose to treat "species-specific or generalized" as a question asked separately for each predictor, and to answer it with backward elimination. Each remote-sensing metric enters the model as a restricted cubic spline, meaning a smooth piecewise-cubic curve forced to be straight beyond its outer knots, where data are sparse. The model contains a shared curve per metric plus optional per-species deviation curves. Recursive feature elimination is the procedure that repeatedly refits the model and drops its weakest component. Here it removes whole curve blocks in a fixed order: species deviations first, then shared curvature, then the shared linear term. The end product is a compact, auditable equation that states, metric by metric, where species identity matters. We expect it to match separate species models in accuracy with far fewer parameters, and to extrapolate better to trees larger than any in training.
Core Claim
The gain from species-specific biomass modeling is concentrated in a small number of predictor-by-species curve deviations. A hierarchically pruned spline model can recover that gain while staying close to a pooled model in size. Concretely, let $s(i)$ be the species of tree $i$ and $x_{ij}$ its value on metric $j$. We model
$$\ln \mathrm{AGB}_i = \alpha_{s(i)} + \sum_{j} \Big[ \beta_j x_{ij} + f_j(x_{ij}) + g_{j,s(i)}(x_{ij}) \Big] + \varepsilon_i .$$
Here $\alpha_s$ is a species intercept, $\beta_j x_{ij}$ is the shared linear term, and $f_j$ is the shared nonlinear part of a restricted cubic spline with $k$ knots at fixed percentiles. The term $g_{j,s}$ is a species-specific spline deviation, constrained to sum to zero across species. A block can be removed only after every block nested beneath it has gone. Every elimination step refits all surviving blocks without penalty, so the final curves are ordinary least-squares splines with no shrinkage bias. The step removes the block whose deletion costs least in inner cross-validated log-likelihood, and elimination stops when outer-fold error rises. Predictions on the biomass scale use a smearing correction for log back-transformation.
Relation to Prior Work
Three groups of papers motivate the question. The Beijing LiDAR study and the urban object-oriented study (Aboveground Biomass Inversion Based on Object-Oriented Classification and Pearson–mRMR–Machine Learning Model) both show that stratifying by species or vegetation type helps. The urban study's two strata ended up with entirely different filter-selected feature sets. Both papers make stratification an all-or-nothing choice at the level of the whole model, and both select features with fixed rankings or filters rather than refitting. We make stratification a per-predictor choice. The Beijing study is also a constraint: its generalized additive model performed poorly on noisy point-cloud metrics. Our design answers that with fixed, few-knot, tail-linear splines and explicit pruning, and the answer has to be tested rather than assumed. The plot-level allometry study motivates the hierarchy: if ontogeny dominates, the shared size curve $f_j$ should carry curvature while most $g_{j,s}$ get pruned. That paper uses only log-log power laws, which cannot bend with life stage.
The method draws on two sources. Elimination with refitting comes from the recursive feature elimination tradition, used for example to cut 140 earth-observation predictors to 82 in an invasive-plant classifier (An earth observation and explainable machine learning approach for determining the drivers of invasive species — a water hyacinth case study). That study eliminated single columns and read thresholds off SHAP plots after the fact. We eliminate structured blocks and read thresholds directly from the fitted splines. Selecting groups of spline coefficients jointly comes from penalized work such as Saturating Splines and Feature Selection and Recursive $\ell_1, \infty$ Group Lasso. That work is a direct competitor, and we test against it rather than arguing it away.
Research Program
The first component is the estimator itself. We will implement the nested-block elimination described in the Core Claim. We will run a sensitivity analysis over 3, 4 and 5 knots, and report how often each block is selected across bootstrap resamples, so that the pattern of species-specificity comes with a stability estimate.
The second component is the scientific analysis. For each dataset we will report which metrics keep species deviations and which keep only shared curvature. We will also locate inflection points analytically from second derivatives of the fitted splines. The prediction is that size-related metrics, such as upper height percentiles and crown area, keep shared curvature. Canopy-density metrics such as laser point count should be more likely to keep species deviations, because they act as a proxy for branching architecture and wood density, which remote sensing cannot see directly.
Data and Evaluation
The primary testbed is the Beijing UAV-LiDAR dataset of 303 measured trees of Pinus tabulaeformis and Robinia pseudoacacia. Its features are height percentiles and point counts from trees segmented with normalized cuts. We will request this data from the authors. Its availability is a project risk, and we will report outcomes conditional on access.
The second testbed is ReforesTree (ReforesTree: A Dataset for Estimating Tropical Forest Carbon Stock with Deep Learning and Aerial Imagery), about 4,600 agroforestry trees with species labels and drone RGB imagery. We will use crown geometry and color-index features extracted from matched crowns. Its carbon targets are themselves computed from DBH allometries, so results there measure how well the method recovers an allometric structure, not true biomass. Its imbalance toward cacao and banana makes it the natural test of rare-species borrowing.
The third testbed is the urban dataset of 88 points in two strata, which tests the very small sample regime. The fourth is the Mt. Wuyishan plot data of 9,105 trees and 15 species, used only with DBH and height. It asks whether the method rediscovers the finding that ontogeny dominates phylogeny.
All evaluation uses repeated nested cross-validation, blocked by plot or site so that neighboring trees never straddle train and test. Metrics are RMSE and bias in kilograms, $R^2$, the number of fitted parameters, and per-species RMSE for the rarest third of species.
Baselines
The naive reference point is a pooled log-log power law in DBH or height, the default in current practice. Any proposal must beat it on rare-species error, or species-specificity buys nothing.
The strongest existing method is the per-species random forest or Cubist model of the Beijing study. It reports $R^2 = 0.89$ and RMSE $= 22.87$ kg for species-specific random forest and Cubist on that dataset. We will rerun it on the same folds and the same feature pool, with an equal hyperparameter search budget. We will run it both with the authors' feature ranking and with standard column-wise recursive feature elimination. Beating it means matching its RMSE within fold-to-fold noise using at least an order of magnitude fewer parameters, or strictly beating it on held-out large trees.
On the urban data, the reference numbers are the forest-stratum random forest at $R^2 = 0.77$, RMSE $= 5.21$ kg/m², and the agri-grass XGBoost at $R^2 = 0.86$, RMSE $= 0.23$ kg/m², both with Pearson–mRMR selection. We will rerun these on our folds. On ReforesTree, the only site-level figure in the file is a fine-tuned CNN predicting 7 kg against an 8 kg ground truth at Site 1. No reported tree-level RMSE was found on Lacuna for ReforesTree, so we will train that CNN ourselves as a reference.
The competitor that most threatens the idea is a hierarchical group lasso. It uses the same spline blocks with nested groups, tuned by the same inner cross-validation, and is then refit without penalty on the selected support so that it does not suffer from shrinkage bias.
Three ablations isolate the key ingredients. Column-wise elimination on the raw spline columns shows the damage done by breaking spline structure. Fully separate species splines, with every $g_{j,s}$ kept, test whether pruning matters. Pooled splines with no $g_{j,s}$ test whether species deviations matter at all.
Why not just…?
Why not just use per-species random forests, which already score 0.89? Random forests predict a constant beyond the largest training tree. Large trees hold most of the carbon and are the rarest in samples. The decisive test is to train on trees below the 85th height percentile and test on the rest. If the forest's bias on those tall trees is no worse than the spline model's, this objection wins on accuracy. Our contribution would then shrink to interpretability and parameter count, and we would say so.
Why not just use a group lasso over spline blocks? This is the real threat, because nested group penalties can encode the same hierarchy in a single fit. Our bet is that refit-based elimination is more stable when only hundreds of trees are available. We will measure that with bootstrap selection frequencies and with error on rare species. If the refit lasso matches us on both, we will adopt it as the estimator. The per-predictor map of species-specificity remains the scientific result either way.
Why not just fit a mixed-effects log-log allometry with species random slopes? This approach already borrows strength across species and is simple to explain to auditors. Its weakness is that it allows species to differ only in intercept and slope, never in curve shape. The Wuyishan analysis settles the question. If pruned splines retain shared curvature in the 5–15 cm DBH range, where the plot study found generalized equations deviate most, and gain accuracy there, the power law is too rigid. If they retain no curvature, the mixed model is enough and we will recommend it.
Decisive Checks
The idea succeeds if three things hold together. The pruned model must keep species deviations for only a minority of metrics, consistently across bootstrap resamples. It must match per-species random forests within noise on the Beijing data. And it must cut tall-tree extrapolation bias relative to them. It also has to improve rarest-species RMSE on ReforesTree compared with fully separate species models. If pruning instead keeps nearly every deviation block, then species really do differ everywhere and the current practice of separate models is vindicated. That is a negative result, and we will report it.
References
- Forest emissions reduction assessment from airborne LiDAR data using multiple machine learning approaches
- Allometric Equations for Aboveground Biomass Estimation in Natural Forest Trees: Generalized or Species-Specific?
- Aboveground Biomass Inversion Based on Object-Oriented Classification and Pearson–mRMR–Machine Learning Model
- An earth observation and explainable machine learning approach for determining the drivers of invasive species — a water hyacinth case study
- Saturating Splines and Feature Selection.
- Recursive ℓ1, ∞ Group Lasso
- ReforesTree: A Dataset for Estimating Tropical Forest Carbon Stock with Deep Learning and Aerial Imagery